In this paper, we propose an innovative channel-sensitive autoencoder (CSAE)-aided end-to-end deep learning (E2EDL) technique for joint geometric probabilistic shaping. The pretrained conditional generative adversarial network (CGAN) is introduced in the CSAE which performs differentiable substitution of the optical fiber channel model under variable input optical power (IOP) levels. This enables the CSAE-aided E2EDL to design optimal joint geometric probabilistic shaping schemes for optical fiber communication systems at varying IOPs. The results of the proposed CSAE-aided E2EDL technique show that for a dual-polarization 64-Gbaud signal with a transmission distance of 5 × 80 km, when the modulation format is a 64-quadrature amplitude modulation (QAM) or a 128-QAM, the maximum generalized mutual information (GMI) level learned via CSAE-aided E2EDL is 5.9826 or 6.8384 bits/symbol under varying IOPs, respectively. In addition, the pretrained CGAN, as a substitution for optical fiber transmission model, accurately characterizes the distortion of signals with different IOPs, with an average bit error ratio (BER) difference of only 1.83%, an average mean square error (MSE) of 0.0041 and an average K-L divergence of 0.0046. In summary, this paper delivers new insights into the application of E2EDL and demonstrates the feasibility of joint geometric probabilistic shaping-based E2EDL for fiber optic communication systems with varying IOPs.

In this work, a low-complexity data-driven characterized-long-short-term-memory (C-LSTM)-aided channel modeling technique is proposed for optical single-mode fiber (SMF) communications. To fully utilize the sequence correlation learning ability of traditional long short-term memory (LSTM) networks and solve the gradient explosion problem, the feature information is introduced into the traditional LSTM input layer to better characterize the intersymbol interference caused by dispersion in SMF modeling. The simulation results show that the proposed C-LSTM can effectively alleviate the gradient explosion problem with a stable and ultimately lower mean square error (MSE) than traditional LSTM. Compared with the split-step Fourier method (SSFM) and the conditional generative adversarial network (CGAN), the proposed C-LSTM has superior computational complexity. Moreover, due to the sequence correlation learning ability inherent to C-LSTM, coupled with the flexibility of feature information selection, the proposed C-LSTM-aided modeling technique has a higher modeling accuracy than traditional LSTM. Moreover, the C-LSTM-aided modeling technique can be effectively extended to other channel modeling applications with strong sequence correlations.

The probabilistic shaping (PS) technique is a key technology for fiber optic communication systems to further approach the Shannon limit. To solve the problem that nonlinear equalizers are ineffective for probabilistic shaping optical communication systems with non-uniform distribution, a distribution alignment convolutional neural network (DACNN)-aided nonlinear equalizer is proposed. The approach calibrates the equalizer using the probabilistic shaping prior distribution, which reduces the training complexity and improves the performance of the equalizer simultaneously. Experimental results show nonlinear equalization of 120 Gb/s PS 64QAM signals in a 375 km transmission scenario. The proposed DACNN equalizer improves the receiver sensitivity by 2.6 dB and 1.1 dB over the Volterra equalizer and convolutional neural network (CNN) equalizer, respectively. Meanwhile, DACNN converges with fewer training epochs than CNN, which provides great potential for mitigating the nonlinear distortion of PS signals in fiber optic communication systems.